v2.0.4

AgentScope Java 2.0.4 adds OpenAI Responses API support, a redesigned AgentScope Service, new channel and storage integrations, and stronger isolation for concurrent agent runs, tools, and skills.

Added

Core / Models

• Add the OpenAI Responses API module (#3078).
• Support returnDirect tool results to short-circuit the reasoning loop (#2891).
• Expose detailed chat usage token metrics and model-fallback failover listeners (#3215, #3145).
• Support Anthropic bearer token authentication (#3038).
• Add a type-safe Jev System One client, Spring Boot starter, and example middlewares (#3240).

Harness / Tools / Storage

• Add per-run execution control with AgentRun / RunControl and explicit runtime context propagation (#3279).
• Add request-scoped tool configuration and isolate tool visibility, streaming callbacks, and session skill resources (#3283).
• Add opt-in sharedLocalWorkspace to filesystem specs (#3247).
• Allow disabling built-in web tools and injecting an HTTP client into them (#3075, #3103).
• Add configurable MCP metadata propagation per server and tool (#3255).
• Add MongoDB storage extension (#2698).
• Automatically prune old E2B native snapshots per sandbox (#2555).

Service / Channels

• Introduce the redesigned AgentScope Service implementation (#3080).
• Add Personal Weixin iLink channel integration (#3184).
• Add DingTalk HTTP callback receive mode alongside Stream mode (#3177).
• Add pluggable channel access-token storage and inbound-event deduplication, with idle eviction for bot-loop guards (#3182, #3175).
• Expose inbound channel media messages as neutral metadata (#3180).

Refactored

Core / Runtime

• Remove the unused message role parameter (#3045).
• Removed activeRc, agent-level getRuntimeContext(), and the deprecated RuntimeContextAware hook setter contract. RuntimeContext is now passed explicitly through middleware and tool parameters (#3279).
• Removed Harness repository context suppliers and the context-less skill-promotion overload. Explicit context is carried through skill lookup, creation, edits, resource operations, archiving, and security rollback; context-less repository operations use the default namespace (#3279).
• Derive Version.VERSION from the Maven project version (#3087).
• Migrate the documentation site to Mintlify while preserving legacy URLs and restoring hosted language switching (#3081, #3082, #3276).

Core / Tools / Skills

• Removed shared skill activation and toolkit-binding APIs in favor of explicit request/session state. Deleted public types: RegisteredSkill, ToolkitAware, agui.model.ToolMergeMode, ToolInjection. Deleted members: setInternalChunkCallback, syncToolGroupStates, isSkillActive, replaceMetaActiveGroups (#3283).
• AG-UI's borrow-mutate-restore ToolInjection model is gone; ToolRequestConfig composes external schemas with the agent registry per invocation (#3283).

Model Providers

• Pass Gemini tool parameter schemas directly as native JSON Schema (#3051).
• HttpTransportConfig#getHttpVersion() now returns null by default instead of HTTP_2. Downstream code calling this getter must add null-check handling to avoid NPE. Set httpVersion(HTTP_2) explicitly to preserve the previous behavior on plaintext HTTP URLs; https:// users are unaffected (#3214).
• DashScopeMessage and OpenAIMessage no longer expose cacheControl (getter, setter, builder method). Message-level cache_control values are no longer emitted or migrated.

Fixed

Core / State / Concurrency

• Isolate concurrent invocation contexts and cancellation of individual running or queued executions (#3279); prevent cross-session tool and skill interference (#3283).
• Fix version-0 state migration CAS conflicts, atomically return JDBC unconditional-write versions, and cache versions after unconditional writes (#3165, #3220, #3236).
• Evict slot versions with the state cache, avoid deserializing the State marker interface, and persist state after empty agent results (#3074, #3048, #3049).
• Preserve failed summary termination state and handle synchronous model-fallback failures (#2757, #3149).
• Emit tool-result events for denied HITL calls and support custom denial reasons (#3104, #2546).
• Make JSON Schema generation thread-safe (#2796).

Model Providers

• Preserve reasoning details from all streaming chunks and thought signatures across JSON persistence round-trips; remove redundant OpenAI tool metadata (#3128, #2910, #2914).
• Move cache control to content blocks (#2878).
• Decouple DashScope streaming from thinking mode, allow custom multimodal endpoint detection, and offload blocking embedding SDK calls (#3137, #3163, #3265).
• Correct Gemini token usage accounting (#3034).
• Serialize Ollama tool-result names as tool_name and expose thinking output as ThinkingBlock (#3093, #3146).

Harness / Sandbox / Integrations

• Preserve non-UTF-8 file uploads and expose session workspace paths in prompts (#2456, #3020).
• Rebind Kubernetes remote snapshots on state resume and preserve wrapping newlines in commands (#3025, #3204).
• Improve Docker sandbox native file transfer and reject truncated downloads (#2923).
• Keep MCP connections open for AgentRun sandboxes that are not owned by the caller; preserve image URLs in MCP tool results (#2302, #3053).
• Allow path separators in JDBC sandbox-shaped slot IDs (#3233).
• Preserve AG-UI tool results after permission resume, assign unique IDs to discontinuous text segments, and clear active-run markers before terminal signals (#3100, #3010, #3109).
• Propagate WeCom send rejections as stream errors (#3173).
• Use modular core and AG-UI dependencies in Spring Boot starters (#3156).
• Normalize Milvus L2 retrieval scores and escape WordReader Markdown table cells (#3070, #3169).

Documentation

• Add bilingual Agent Harness building guides and rebuild AgentScope Service product guides (#3139, #3136).
• Document Jev structured decision integration and correct the documented CompactionConfig.keepTokens default (#3274, #3144).

Full changelog: v2.0.3...v2.0.4